CAS-Spawned ScienceClaw Launches AutoProject: AI Research Agents Move From Tasks to Whole Projects

量子位 · wechat · 2026-08-24

ScienceClaw, the research agent from CAS-incubated Zhongke Zidong Taichu, has unveiled AutoProject, a project-level autonomous research engine that moves AI from executing isolated tasks to owning full scientific projects: given only a research goal, the system handles planning, task decomposition, long-horizon execution, evidence verification, dynamic repair and result consolidation.

The engine has three core layers: Project2Task generates executable task networks with flexible topologies; TaskExecutor sustains long-horizon autonomous execution with anomaly handling and experiment re-runs; and EviGraph, an evidence-graph verification system, checks cross-task logical consistency, scoring 0.865 on ARCBenchML versus a 0.596 best baseline, with ClaimSupportRate up 40.74%. The team stresses "autonomous research is not unmanned research" — researchers can intervene at any point. Built on the Zidong Taichu multimodal LLM and a hierarchical multi-agent architecture, ScienceClaw covers life sciences, materials, chemistry, physics and astronomy.

Original post →

More from Apps

Apps channel →